jorisvandenbossche commented on pull request #11818: URL: https://github.com/apache/arrow/pull/11818#issuecomment-1000236783
> At the moment some of the code (ns, us, ms, s, m, h, d, w) is UTC based even when timezones are present. UTC always converts to a local time so nonexistent is not an issue while ambiguous could be. In pandas we always round in local time (some discussion about this in https://github.com/pandas-dev/pandas/issues/37592), and then you can actually run into nonexistent times as well. The reason we do this (with its limitations ..), is because not every timezone has an exactly rounded / hourly offset, and so because of that rounding can give wrong results. For example, in Australia in summer there is a UTC+10:30 timezone: ```python >>> ts = pd.Timestamp("2021-07-01 12:17:00", tz="Australia/Lord_Howe") >>>ts Timestamp('2021-07-01 12:17:00+1030', tz='Australia/Lord_Howe') >>> ts.round("H") Timestamp('2021-07-01 12:00:00+1030', tz='Australia/Lord_Howe') # with this PR with pyarrow >>> arr = pa.array([ts]) >>> pc.round_temporal(arr, unit="hour") <pyarrow.lib.TimestampArray object at 0x7f743c72dc40> [ 2021-07-01 02:00:00.000000 ] >>> pc.round_temporal(arr, unit="hour")[0] <pyarrow.TimestampScalar: datetime.datetime(2021, 7, 1, 12, 30, tzinfo=<DstTzInfo 'Australia/Lord_Howe' +1030+10:30:00 STD>)> ``` So with pyarrow that actually gives the "wrong" value now because of relying on rounding in UTC. As I mentioned in the above-linked pandas issue, rounding in local time _does_ have some serious drawbacks (eg those annoying nonexistent times that actually should be possible to round properly without error), so it would be nice if we could somehow detect the cases where it is fine to do this. But doing that robustly might be quite complex ... -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
